Aiming at the problems of low accuracy, easy to overfitting and poor robustness caused by traditional prediction algorithms in high-dimensional non-balanced small data sets, this paper supposed an approach of predicting nursing turnover based on gradient-enhanced ensemble classifier CatBoost, processing category features, optimizing parameters with BOHB (Bayesian Optimization and Hyperband) and cross validation. Finally, the paper applies the algorithm in nursing turnover data of Shanghai public hospitals with high-dimensional non-balanced small data sets, compared with common algorithms such as XGBoost, Random Forest, Support Vector Machine. The experimental results show that the model is better in accuracy and has good robustness. This proposed model is a promising alternative for prediction of nursing turnover in the hospital human resources management.